Sep 1 – Oct 1, 2026 · 6 items
The AI week, for a product manager
A real edition, written for: Product Manager
The one thing
Which of your roadmap assumptions would break if NVIDIA changed Hugging Face’s terms tomorrow?
NVIDIA to acquire Hugging Face
NVIDIA has agreed to acquire Hugging Face, the open model sharing platform. Chief executive Jensen Huang announced the deal on NVIDIA's blog. He put the price at just under $13 billion. NVIDIA says Hugging Face will stay open to the whole AI industry. Developers will keep their choice of models, frameworks, clouds and chips. NVIDIA compute will not be required to build or deploy through the platform. Hugging Face will continue to support open source and open weight models from all builders. NVIDIA says its infrastructure and engineering can improve reliability, safety, model evaluation and deployment. The Hugging Face team and its brand will remain. No closing date or regulatory timeline was given.
Why it matters for you
Your SaaS roadmap depends on open models and their ecosystem. That ecosystem just got a new owner.
- NVIDIA pledged to keep Hugging Face open to all builders and clouds
- Your vendor choices may narrow if NVIDIA ties the platform to its stack
- The deal values Hugging Face at nearly $13B
Try this
Review your model and hosting dependencies against NVIDIA’s stated commitments
Sources
OpenAI DevDay: Dots and faster models
OpenAI announced over twenty updates at DevDay 2026. It introduced Dots as always-on agents for ongoing tasks. GPT-6.1 Sol launched with near-Astra performance at one-fifth the token cost. Ultrafast mode offers up to eight times faster token generation in Codex. Private Intelligence now supports zero data retention with private safety processing. Codex can now run in the cloud or via a refreshed CLI with voice control. Code Review in ChatGPT desktop app enables automated and manual reviews. Codex Security Cloud scans repositories and prepares fixes in the cloud. Decisions API entered limited preview for real-time classification. Agents API now supports computer use for software interaction. Bedrock Managed Agents brings OpenAI agents to AWS. Plugin extensions allow sidebar integration in ChatGPT. Improved plugin creation and discovery tools were released.
Why it matters for you
Dots are always-on agents that could draft your PRDs or specs while you sleep.
- Dots run 24/7 on tasks and message progress in ChatGPT or Slack
- GPT-6.1 Sol cuts token costs to one-fifth of Astra
- Ultrafast mode in Codex speeds up coding tasks
Try this
Test a Dot on a low-stakes spec draft this week
Paste this into your AI tool
Context: I’m a Product Manager drafting a PRD for a B2B SaaS feature Task: Draft the PRD outline based on [paste user research notes] Constraints: - Use the same headings as our last PRD - Flag any gaps in the research Output: A PRD outline with clear next steps
Sources
Jev cuts decision costs by 99%
TypeSafe released Jev, a model that makes typed if-then decisions without generating text. Theodore Lee released SemIf, an open-source version with the same approach. Both cut inference costs by 99% for basic programming logic. They run attention math once and skip full token generation. Jev costs $0.042 per million input tokens with no output cost. On a 2,000-input-token call, Jev is 82 to 209 times cheaper than frontier models. In tests, Jev and SemIf nearly doubled classification accuracy on 98 email threads. They also handled 8 of 31 live emails without errors. The approach suggests other programming primitives may see similar specialization. It also splits AI economics into discovery and production use cases.
Why it matters for you
Your roadmap prioritisation hinges on quick, cheap decisions. Jev makes them both.
- Jev runs typed if-then decisions without generating text
- Costs drop from dollars to fractions of a cent per call
- Accuracy nearly doubles on real email threads
Try this
Pilot Jev for one prioritisation decision this sprint
Paste this into your AI tool
Context: I need to prioritise features based on [paste user research insights] Task: Rank these features by impact and effort using Jev Constraints: - Use only the data provided - Output a ranked list with a one-line rationale per item Output: Ranked feature list with rationale
ElevenLabs v4: 90 languages, 10s cloning
ElevenLabs released Eleven v4 and Eleven v4 Turbo text-to-speech models. Eleven v4 uses a new architecture for more emotive, controllable speech. It supports 90+ languages and 17,500+ voices. Professional Voice Clones are supported again after being absent in v3. Audio tags like laughs or whispers are followed more reliably. Context stitching keeps long-form audio consistent. Regenerations retain the same speaker identity. Eleven v4 Turbo is a low-latency variant for real-time use. It has ~100 ms median inference latency and ~150 ms time to first speech. Both models are available via apps, API, and agents. They support bidirectional streaming for agent loops. Professional Voice Clones work identically across both models. The models are available now.
Why it matters for you
User research often includes voice feedback. v4 makes it easier to analyse at scale.
- Voice cloning now needs only 10 seconds of audio
- Supports 90 languages, including Swiss German
- Better handling of long-form audio and expressions
Try this
Transcribe and analyse one user interview with v4 this week
Paste this into your AI tool
Context: I have a user interview in [language] about [product area] Task: Transcribe and summarise key insights from [paste audio file link or text] Constraints: - Preserve speaker identity and tone - Highlight pain points and feature requests Output: A structured summary with quotes
Manus 2.0: agents with Cloud Computers
Manus launched Manus 2.0 with a new agent harness called Cascade. Cascade reduces token use, task time and cost versus the prior system. Cloud Computer offers dedicated environments for projects needing persistent hosting. Automations now start from external service events as well as schedules. Manus Studio replaces the desktop app as a shared workspace for people and AI. Studio adds Video Editor for timeline-based video editing and fine-tuning. Video Editor includes Alchemy mode for higher-quality creative direction. Game Dev enables building, editing and publishing multiplayer games. Game Dev supports one-click multiplayer setup via Cloud Computer. Remote Control and Computer Use let users command their computer from elsewhere. Cue is a new standalone app for personal agents on phone and desktop. Cue agents can have their own email, phone, wallet and computer. Cue agents can work in group chats to hand tasks to each other.
Why it matters for you
Your specs often need persistent environments. Manus 2.0 offers them for agents.
- Cloud Computer provides dedicated environments for long-running tasks
- Automations trigger from external events like Slack or GitHub
- Cue app gives agents their own email, phone, and wallet
Try this
Map one workflow that could run unattended in Manus 2.0
Sources
Perplexity Photon: faster, cheaper search
Perplexity built Photon, a new in-house retrieval and ranking engine. It replaces an adapted open-source system that hit cost, latency, and recovery limits. Photon uses compact data formats, batched reads, and separate index building from query serving. It now powers Perplexity’s search pipeline. A new fast preset for the Search API offers 160 ms p50 and 230 ms p95 latency. Across six benchmarks, the fast preset cut estimated model-plus-search cost per task by 68%. Task quality remained comparable to the default preset. The fast preset is available to developers today.
Why it matters for you
Discovery relies on quick, accurate search. Photon cuts both latency and cost.
- Fast preset offers 160 ms p50 latency
- 68% lower estimated cost per task
- Task quality remains comparable to default
Try this
Run one discovery query through Photon’s fast preset
Paste this into your AI tool
Context: I’m researching [specific user problem or market gap] Task: Find and summarise the most relevant recent insights Constraints: - Use only public, reputable sources - Prioritise data from the last 12 months Output: A bullet-point summary with source links
Build this
Every week, one small thing to build with AI in something you actually care about. No work in it. Five minutes to set up, and worth keeping if it earns a second run.
A meal plan for the week, with the worst dish argued into the top spot.
Five minutes to set up
Context: I’m planning meals for the week using [list ingredients already at home] Task: Propose a 5-day meal plan Constraints: - Use only the ingredients listed - Each meal must serve 2 people Output: A simple meal plan with one dish per day
- List the ingredients you have at home in one line
- Run the prompt in ChatGPT to get the first meal plan
- Rank the dishes from best to worst
- Rewrite the plan to argue the worst dish is actually the best
Yours arrives Thursday.
This one was written for a product manager. Tell us what you do and the next one is written for you — same news, your job, once a week.